The Smallest Sovereign

Sep 1, 2026

The citizen-scale companion to When the Models Go Dark: it begins in one person’s archive and ends in national strategy.

Twenty-four seconds

This month, a scoring engine I built inside my private system, which I call the perceptron, ranked 22,173 people in twenty-four seconds. Go players have a name for what that moment felt like: move 37, AlphaGo’s one-in-ten-thousand move against Lee Sedol in 2016 — shorthand for the instant you see, from inside your own game, that the regime has changed.

The archive it ran against is a system I have been building for years: my private Memex, named after Vannevar Bush’s proposal. I am now generalizing parts of that architecture in the open-source MemexLab Engine, still an early preview. The Memex is not a folder of notes but a living knowledge graph. It now holds 30,505 Markdown pages: 22,173 people, 3,530 companies, and 1,904 analyses of books, reports, and decks. More than 48,000 typed links join them, and some 600 curated source PDFs sit behind the analysis pages. The graph expands a little every day I live and work. Everything is reachable in seconds. The friction has not vanished; it has moved from retrieval to maintenance. A task can draw on all or part of the archive, using a local model or an API. I asked it who to approach for a partnerships push, in what order and why: a task that would once have taken a week.

The ranking showed me what the system could do. What it did not yet know showed me what it could become.

The path was precise: thousands of names waiting to be linked to the right company pages, probable duplicates already flagged and excluded, and one missing layer of evidence — who had actually interacted with whom. The model had finished in twenty-four seconds. The distance between a fast answer and one I could act on did not require a better model. It required a better graph on my side of the API: name resolution, deduplication, and sharper rules for what counts as evidence. That work compounds. Next month’s ranking will be better even if the model does not change at all.

That experience comes down to three sentences, and they are the whole of this essay:

The library is the asset. The librarian is the engineering. The model is the replaceable input.

What follows works out what those sentences mean for a person and then for a state. The question of who owns intelligence runs all the way up.

The room, not the sky

In a talk given to founders this year, Garry Tan, the president and CEO of Y Combinator, put a frame around what I had been living inside. “Everyone is watching the sky,” he said, “and the thing they’re watching for is already in the room.”

The thing being watched for is artificial general intelligence, expected to arrive as an event: an announcement, a benchmark, a day the sky changes colour. Tan’s argument is that this is a category error. It is arriving diffused: as an agent running on your infrastructure, reading your files, finishing a job while you sleep. He calls this personal AGI: general intelligence for one person — you.

Whether it deserves the name AGI is a definitional argument nobody wins. What is beyond argument is that this is a different operating regime: persistent memory, encoded procedures, tools, and a replaceable model compounding together. It is not a chatbot subscription. In Tan’s phrase, that is “a corporate AGI you don’t own.” Assistants have memory now; the difference is custody.

A chatbot’s memory of you lives in the vendor’s store and format. It is not generally exportable as a portable, model-ready corpus. You cannot simply take it with you, point another model at it, or read it whole. When the company pivots, your context pivots with it. An owned system starts with a corpus you hold, in a format you can move, that any model can read. It improves every day you use it. Very few people yet operate a genuinely portable version.

The equation Tan offers for the next decade is worth writing down precisely:

A frontier model — rented, cheaper every quarter

+ your context — owned, unique, ideally possessed by no one else on earth

+ a harness that wires them together

= an agent that acts like a very fast version of you.

Frontier capability is rented. Accumulated context can remain yours. Running my own system daily has taught me a sharper version of the same law: you can outsource your thinking. You cannot outsource your understanding. The system does not understand for you; it preserves what you have read and decided, and hands that context back at the moment of decision.

There is one important qualification. A fixed level of capability commoditizes: yesterday’s frontier gets cheap while today’s remains scarce. Context determines who turns either into durable capability. Replaceability is a design rule. A system built to survive a model swap gets better every time the market does; one welded to a single vendor inherits that vendor’s fate.

The test has already run. Tan says he sees 2x people and 100x people using the same Claude: same weights, same price, enormous variance. The difference sits above the weights, in context, workflow, skill, problem selection, and whether the right information arrives at the right step. Skill and judgment remain personal. Architecture can preserve, port, and compound context. That makes it the layer strategy can act on. The leverage that is yours was never in the model alone.

From a point to a field

Tan borrows a useful frame from Spinoza. The part I need from Spinoza is structural, not theological. He shifts the frame from a single external centre to relations distributed across the whole: from a point to a field.

The public imagination still treats AGI as a single intelligence in a data centre, owned by a particular company and arriving on a particular date. But useful machine intelligence is already appearing in a more distributed form: as infrastructure. A terminal window. A folder of Markdown files. A job that finishes overnight. This does not make today’s tools AGI. It means capability is spreading through people, tools, and institutions before anyone can point to a single arrival event.

Spinoza also offers conatus: the tendency to persist and increase one’s capacity to act. In his vocabulary, joy is what such an increase feels like. When an agent compresses a week of work into an afternoon, the result feels like more than convenience. It feels like more room to act.

A nation is not a person. But it too depends on its capacity to preserve what it knows and extend what it can do.

The arithmetic, honestly stated

Tan reports that in 2013, building a product at night as a YC partner, he shipped roughly fourteen useful lines of code a day. This year, while running YC full-time, he calculates his output at roughly 400 times that amount. He discounts the number himself. After punishing it for verbosity and self-flattery, his lower estimate is still 8x. I would go one step further: 8x is an anecdote, not an empirical floor. I keep it on the table because the phenomenon it points to is now too common to dismiss. Policy built on the deflated end of these anecdotes does not need the headline to be true.

There is one broader data point, though it comes from YC itself: roughly a quarter of its Winter 2025 batch had codebases that were 95% AI-generated. It is a correlation from an interested party, but the narrow reading is enough. The fastest-growing founders in that ecosystem treat AI as a workforce, not autocomplete. Code is simply where the multiplier is easiest to count. The same structure applies wherever the work involves reading a large amount of context and producing a judgment or an artefact.

Seven digits and a library

Every one of these systems runs into a limit that institutions have spent centuries working around: human working memory.

George Miller’s 1956 paper put the span of immediate memory at seven items, plus or minus two; Nelson Cowan’s 2001 reconsideration argues for four. Either way: single digits. The leap from working memory to org charts is an analogy from Tan, extended here; neither paper makes it. But it is a productive analogy. Read this way, every checklist, filing cabinet, standup meeting, briefing note, and chain of command is a prosthetic for a seven-digit brain. The bureaucratic form of the modern state is a compression algorithm for a working memory of about four things.

An agent can hold a million tokens: roughly a thousand pages in Tan’s illustration. The figure is model-dependent, and holding a claim is not the same as synthesizing it well. Still, the order of magnitude creates a different operating regime. Most ministries, regulators, and central banks run procedures designed for the older one.

Now run the number the other way. A thousand pages is a great deal, and it is almost nothing: your life is not three books but a library — every email, every meeting, every decision and the reasoning behind it. A ministry’s life is a national archive; a country’s, a civilization. So the question that decides whether an agent is a genius or a goldfish is:

Who — or what — decides which three books are open on the desk?

That is the librarian problem, and it is what my perceptron run measured to the page. It is not a model problem; more frontier capability does not, by itself, solve it, because it is a question about your corpus, your retrieval, your judgment of relevance.

The pattern has a lineage. Vannevar Bush imagined the memex in 1945: a private mechanized library, an “enlarged intimate supplement” to memory. Andrej Karpathy formalized the modern version this spring as a persistent, LLM-maintained Markdown wiki that he traces straight back to Bush. Tan industrialized it for himself in GBrain, with roughly 220,000 pages and twenty-five years diarized (his figures, self-reported). My Memex is the same species, smaller but a full graph rather than a pile of pages. Its logs show the part the keynotes skip: accumulating the library takes years; making it answerable is where the engineering lives.

That fact should sit beside every serious AI strategy.

The turn: a nation, one scale up

A person and a state are not the same kind of thing, and the analogy should not be pushed past what it carries. What carries is narrow and sufficient. Three dependencies recur at both scales: owned context, reliable retrieval, and continuity across model changes.

A nation that runs on a rented frontier model with no owned context has a corporate AGI it does not own. It resets when the vendor pivots. It knows what everyone else knows. When the company behind it changes its terms, jurisdiction, or geopolitical alignment, the nation’s cognitive infrastructure has to adapt on someone else’s schedule.

This is not a hypothetical failure mode. In June, two frontier models went dark for every customer, everywhere, at once after an external regulatory order beyond most customers’ control. That cutoff opens When the Models Go Dark. Reliance on a revocable API is the default risk when national AI strategy is treated primarily as procurement.

Most sovereign AI programmes understandably begin with compute and models. These are the layers where scale favours the largest economies: export controls matter, capital requirements are high, and incumbents hold a long lead. The opportunity is to give equal strategic weight to data, deployment, and talent, the national forms of library, procedure, and librarian.

The scarce resource keeps moving: first labelled data, then raw text, then GPUs, and now electricity. A fixed level of model capability rarely stays scarce for long. It diffuses and becomes cheaper even as the frontier advances. What resists copying sits around the models: grid connections, institutional memory, and legitimacy. Those assets are built locally and over time.

What Türkiye is getting right, and the next layer

Türkiye’s Artificial Intelligence Action Plan for 2026–2030 is a serious document: ten billion dollars in mobilized investment; a gigawatt of data centre capacity by 2030; domestic model work through TÜBİTAK’s Bilge, the T3–Baykar programme, and HAVELSAN’s MAIN platform; 10,000 advanced specialists and 100,000 practitioners; AI literacy for five million citizens across all 81 provinces; at least 2,000 public datasets in a National Data Library. Read through the lens of this essay, three things stand out.

The National Data Library is the right instinct. It is not a rival to the gigawatt but its natural complement: the gigawatt supplies computation, and the Library is what makes that computation about Türkiye. In the strategies I have read, the maintained corpus rarely appears as an appreciating asset. The literacy programme has similar promise. Five million people is a serious ambition, and personal AGI is fundamentally a distributed phenomenon.

The next step is to build the librarian. Two thousand curated datasets is a considerable achievement in data governance and a strong foundation for context engineering. Datasets are records; context adds situated, procedural, decision-linked knowledge, including the reasoning behind a decision rather than only its result. The Library becomes more valuable when those datasets are paired with retrieval, procedure corpora, and agent-queryable institutional memory. That layer can be built from day one, with the plan’s own infrastructure as its natural home.

A larger opportunity is not yet explicit in the plan: personal AI sovereignty, the citizen’s own compounding intelligence asset. Existing data-portability work is useful, but it usually moves records between institutions. It does not yet treat a citizen’s accumulated context as a productive asset to be built. Türkiye’s literacy and data programmes provide a strong base from which to explore that position.

The high-leverage hypothesis worth testing

The economics are promising, but they remain hypothetical.

A gigawatt of AI-dedicated compute is capital expenditure with a depreciation schedule, a foreign supply chain, and significant external dependencies. Some of it is necessary. Energy and a minimum inference reserve are the layers whose absence is fatal in a cutoff. But for a middle power, this is a field where scale favours larger economies. Its strategic value lies primarily in continuity, not in matching the frontier dollar for dollar.

Personal AGI has a different profile. It is an asset that appreciates with maintenance: a corpus kept current, resolved, and linked is worth more every year; neglected, it decays into a pile of files. That upkeep is the librarian’s work, not magic, and it is exactly where my perceptron run showed the next gains would come from. It is also distributed. No single facility has to carry the whole system. That resilience is not absolute, which is why the custody and security architecture below is a precondition rather than a garnish. The expensive input is the part getting cheaper by the quarter. What remains — storage, security, inference, data hygiene, and maintenance — is a cost a person or small institution can actually carry.

Then there is the multiplication. The honest unit is not a multiplier but an hour. Knowledge workers spend a large share of each week reconstructing context: searching mail, re-reading documents, rebuilding the reasoning behind decisions already made. Suppose an owned, compounding system returns five hours a week, a fraction of the headline claims. That comes to roughly six working weeks per person per year. Across a million knowledge workers, it is on the order of a hundred thousand person-years of skilled capacity annually — recovered, not hired.

Türkiye has a natural place to test the hypothesis: 13,000 technology startups in technoparks, 1,700 R&D centres, and more than 246,000 ICT professionals, seventy percent of them under thirty-five. That talent shows up abroad too: Turkish-founded diaspora startups raised $1.1 billion in 2025 and produced three new unicorns. Capital can buy compute and attract talent. It cannot import an ecosystem’s accumulated context fully formed; that grows through the daily work of its own people and institutions.

This is still an illustrative thought experiment, not a measured result. Hours returned matter only if quality holds, which is why the pilot below measures quality-adjusted hours rather than raw speed. Task-level gains do not automatically become national productivity, and such effects often take years to appear in aggregate statistics.

Even after that discount, the asymmetry remains: a middle power does not need to match hyperscaler spending dollar for dollar. It can build an advantage in context that scale alone cannot quickly reproduce. Context is still a relatively open field.

Four objections, taken seriously

An argument this convenient for a middle power deserves adversarial pressure.

First: context sovereignty is incomplete without model continuity. If model access is cut, the harness stops and the library sits unread. Correct. Rent the frontier with a tested exit, archive and test more than one open-weight lineage, hold a minimum inference reserve, and drill the migration. Continuity is insurance; context is the compounding asset. You need both.

Second: machine-readable personal histories create a high-value privacy and security surface. Centralized custody or weak access rules would expose citizens, institutions, and public trust. This is the hardest objection, in any country. Technology alone cannot decide who holds the asset, who sets the rules, and who carries the risk. Governance has to precede diffusion. National capability is stronger when personal context is protected by design.

Third: unequal access could widen productivity gaps. The difference between 2x and 100x users can amplify existing differences in skill and access. That is an argument for making the core capabilities — custody, portability, tools, and literacy — broadly available, following the same logic that widened access to literacy, electricity, and telephony.

Fourth: the evidence remains thin. The 400x is a self-report; the 8x comes from the same witness; the YC statistic is a correlation from an interested party. None is enough to justify scaling on faith. Start from the deflated end of the anecdotes, test the claim in a controlled pilot, measure quality-adjusted hours, and expand only if the results hold.

Five moves

If the analysis holds, the implications are concrete. None of these requires a gigawatt.

1. Build the protocol alongside the models. Türkiye’s domestic model programmes are worth funding, and the plan funds them. Their natural complement costs a fraction as much: an open memory format, a portability standard, a reference implementation, model-swap conformance tests, key-custody standards, and an independent evaluation suite. A shared protocol allows Turkish tools and providers to compete without concentrating dependency in one product, and helps turn the National Data Library from a static archive into an operational asset. The public sector builds the rails; the market builds the trains.

2. Give citizens custody of their context. This cannot mean absolute ownership, because personal context often contains the rights and confidences of other people. What can be established is practical custody: encryption by default, keys held by the citizen, portable storage, user-controlled recovery, and clear rules for exceptional access. Public institutions define interoperability and safeguards; citizen-controlled tools or certified providers operate the stores; independent audits verify the separation of roles. An asset that cannot be securely held will not be built.

3. Aim AI literacy at the skill that compounds. Five million citizens is the right ambition, and reach at that scale is a strategic asset few countries can match. Tool-specific prompting ages quickly; building and maintaining a personal knowledge system compounds across tool generations. Each participant should begin by importing what they already hold — documents, records, and correspondence — into a store they control and grow through daily work. The first week’s exercise is retrieval from one’s own past, not data entry.

4. Give every institution an owned memory and a librarian. Ministries, courts, hospitals, and directorates all face the same constraint: capable people working across fragmented records and procedures. What an institution needs is not a “personal AGI,” but a machine-readable, agent-queryable memory that preserves both decisions and the reasoning behind them, subject to appropriate provenance, privilege, retention, and access controls. Measure memory quality, source traceability, and portability alongside deployment speed.

5. Run the experiment, then measure what compounds. Begin with a controlled pilot across three occupations, a six-month comparison group, and one deliberately boring headline metric: quality-adjusted hours returned per week — an hour counts only if blind-scored output quality holds. Track retrieval accuracy, source traceability, model-swap time, repeated errors, and privacy incidents alongside it.

Then measure durability, not enrolment. The citizen-scale counterpart to the continuity tests in When the Models Go Dark is a drill: how many participants can export their corpus, restore it on infrastructure of their choosing, and re-point it at a different model? That measures whether the asset can survive the loss of any single vendor or external API. Scale only if the results hold.

The lens grinder

Spinoza ground lenses — instruments that extended the reach of the eye. A personal knowledge system should do the same for judgment: not replace it, but carry it farther than memory alone can.

Because this essay began in one person’s archive, it should end there, on a Monday morning. You do not start from zero: you already hold the raw library — years of mail, documents, notes, and decisions. Open a folder. Plain text. When you finish a piece of work, record the decision and the reasoning behind it — the part most archives lose first. Ask a model what you decided about something last year and why; your own past reasoning comes back into view. Link people to companies to decisions. Point this year’s model at the corpus, then next year’s model at the same corpus. The models will change. The accumulated context will remain. Compounding is quiet at the beginning; twenty-four seconds is what it sounds like later.

The same principle scales. After a cutoff, a country can still use the context it has preserved, through the retrieval systems it has built, on infrastructure it can continue to operate. So can a person.

Ownership here means something practical: custody and control of the context on which intelligence compounds.

The smallest sovereign is a person who owns their intelligence — and can carry it forward.

A sovereign nation is built when enough people and institutions can do the same.

Sources and notes

• Garry Tan, “Own Your Intelligence,” Y Combinator Startup School 2026 (video). Parts of this essay reconstruct and extend Tan’s framing rather than merely quoting it: the personal AGI definition, the model + context + harness equation, the Spinoza and conatus framing, the 400x/8x self-estimates, the 2x/100x comparison, the seven-item/million-token contrast, the library/librarian distinction, and the GBrain description originate in the talk. The productivity and corpus figures are self-reported and unverified. The open-source GBrain repository, accessed 31 August 2026, describes a roughly 150,000-page brain, below the talk’s 220,000-page figure; both figures are Tan’s own.

• Andrej Karpathy, “llm-wiki,” 4 April 2026 — a persistent, LLM-maintained, interlinked Markdown wiki as a compounding artefact between immutable sources and retrieval, explicitly related to Vannevar Bush’s memex.

• Vannevar Bush, “As We May Think,” The Atlantic, July 1945 — the memex as a private mechanized library and an “enlarged intimate supplement” to memory.

• Move 37: game two of the AlphaGo–Lee Sedol match, March 2016. Google DeepMind’s account describes it as a pivotal and creative move with a 1-in-10,000 probability of being played.

• Baruch Spinoza: Ethics III, propositions 6–7, on conatus; IIIp11 and its scholium, on joy and sadness as transitions in the power of acting. On Spinoza’s life and work as a lens grinder, see Steven Nadler, Spinoza: A Life, 2nd ed. (2018).

• George A. Miller, “The Magical Number Seven, Plus or Minus Two,” Psychological Review 63, no. 2 (1956); Nelson Cowan, “The Magical Number 4 in Short-Term Memory,” Behavioral and Brain Sciences 24, no. 1 (2001). Neither paper supports treating institutional structure as a consequence of working-memory capacity; that extension is an analogy from Tan’s talk, developed here.

• Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” 12 June 2026; “Redeploying Fable 5,” 30 June 2026. An export-control directive targeting foreign-national access led Anthropic to suspend both models for all users because nationality could not be verified in real time. The controls were lifted on 30 June and access began returning on 1 July. This is the cutoff treated at length in When the Models Go Dark.

• Türkiye Yapay Zekâ Eylem Planı (2026–2030): put into effect by Presidential Circular 2026/9, Resmî Gazete No. 33344, 18 August 2026; announced by the Ministry of Industry and Technology. Plan figures are also reported by Daily Sabah and Türkiye Today. The 13,000 figure refers to technology startups inside technoparks, not technology companies nationally.

• Y Combinator Winter 2025 batch, with roughly 25 percent of companies reporting codebases that were approximately 95 percent AI-generated: TechCrunch, 6 March 2025.

• Sam Altman, “Three Observations,” February 2025 — supporting, not independent, evidence that the cost of using a fixed level of AI capability falls rapidly while frontier capability continues to scale with training and inference resources.

• ICT workforce figures: Republic of Türkiye Investment Office, ICT sector overview — more than 246,000 ICT-sector professionals, 70 percent of them under thirty-five; accessed 31 August 2026.

• Turkish-founded diaspora startups raised $1.1 billion across 41 deals in 2025, with Airalo, Fal, and Periodic Labs reaching unicorn status: startups.watch, “Turkish Startup Ecosystem — Year in Review 2025,” hosted by the Turcorn programme.

• The Memex figures in this essay — the page, person, company, analysis, PDF, and link counts — are from a 31 August 2026 scan of the underlying files and graph index. The latest perceptron scoring run covered 22,173 person pages, with the coverage gaps described above. The hours arithmetic is an illustrative thought experiment, stated as such in the text.

• Disclosure: the private system described in this essay has been operating for years. The author is now building MemexLab, an early-preview product that generalizes parts of that architecture. MemexLab is not the private system described here, and the essay’s policy argument favours open protocols and competition among multiple implementations rather than any single product.